Cornell 2022 Future of Finance Conference
@Fintelligence Academy
9/16/2022, @NYC
Figure 0.1: View from Cornell Tech Manhattan Campus, Roosevelt Island, NYC
This block presents the first-hand summary from Cornell Financial Engineering Manhattan & Rebellion Research’s 2022 Future of Finance conference (link). The conference is hosted at September 16, 2022 in Cornell Tech Manhattan campus, which is located on the scenic Roosevelt Island in New York City.
This event is in celebration of Cornell Financial Engineering Manhattan (CFEM)’s 15th anniversary. A broad range of experts in quant finance, from senior academics to seasoned Wall Streeters and Fintech founders, are invited to join the conference panels to exchange ideas and share opinions.
CFEM, established in 2007 as a satellite New York City campus for Cornell Financial Engineering programs, serves to connect Cornell students with alumni and other practitioners working in the field of quantitative finance.
1 “TL;DR”: Conference Spotlights
Here are the main 💡takeaways from the conference:
Trend of Fundamental Shift in 2022’s Financial Market: quant becomes more discretionary; fundamental becomes more quant.
Financial Machine Learning needs to be nuanced. Financial data are the least stationary of lowest noise-to-signal ratio, where plug-in MLs do not work.
Causality beyond Correlation. Causal discovery application to find the real driving factors for quant models is the new evolving trend as of 2022.
Regime-switching model does not work in real market. Diversified signal/strategies can work to hedge out the uncertain regimes.
Low R-squared for Quant Models. Markets go a lot further than prediction. Be alert of the underlying factors that can not be well captured by quant models, which are the de facto major drivers in market.
What works in real trading - evidence from Numerai. XGBoost and Self-supervised Learning for building feature before training the model to forecast alpha. Linear regression performs relatively poorly in Numerai platform, yet not losing money with low variance.
In the forthcoming sections, we introduce the key nine panels selected from the conference in a chronicle order. While there may be some controversial points of views, we strive to faithfully keep the original points from the speakers, and leave to readers to digest and delve into using their own discretion.
2 Panel #1. The Investment Process
Quantitative resarch is not natural science like Physics
\(-\) Moderator: David Nelson, Chief Strategist & Economist, Belpointe Asset Management & Fmr Financial News Anchor, NewsMax
Danger/Shortcomings of quant models (AI, ML) in market investment
QR (quant researchers), faces an incomplete information in nature science, all data are generated by human. Model may fail in critical way.
It’s easy to conduct close experiments and study variables by isolation. However, Markets can’t be studied in the same way. So, quants face a far more complicated subject than natural science.
Markets is full of human intentions such as fear, momentum, human behavior, leading to its great complexity. Quantitative methods can work but models can fail in very critical way by oversimplification.
Rule of thumb is to never disagree with fundamental managers, Fundamental analysis is more in tune with real-time data and don’t need to deal with historical TS data.
Fundamental managers are more forward-looking versus quant managers. Continue to refine model is going to be the winner.
ESG
Hard to assess if we are on the right track: e.g. nuclear is one way to solve energy problem, but nobody wants nuclear power plant in their backyard.
Path forward is nuclear, with legal support.
There are only finite ways we can solve these (environmental) challenges
Advice for PM - Rationale
Investment process is about having a rationale for decision making. Identifying decision and trading ideas. Threshold/size for positioning using quant methods. Having a reason and being able to define a reason.
Fundamental managers are not good at idenifying size.
The scarce resource is time, given the large-scale data.
Trading Derivatives
- Derivative volume is big. ~$12.4 trillion (unverified approximation)
Sequential set-up for investment process:
Investment ideas
Translate ideas into actions on investment vehicles, where underlying assets are highly correlated
How to enter and size the position.
Risk management and alpha generation
- PMs need to keep in mind that alpha comes with risk management
Current regime shift for investment process
There is big regime change right now, market needs to adapt.
Green. - Resilient grid with deployment of power vehicle. Hydrogen still several years ahead.
Bank of Japan can have massive impact with the yield curve control. It is the only one still doing quantitative easing. Dollar/Yen currency is appreciating at an alarming rate.
Next stage for financial model in the big data era
Reconciliation between ML/AI and Econometrics/Statistics, such as option pricing - gap between quant and data science. Casual inference is required, but it is complex and at a nascent stage.
DS is designed to forecast/prediction. Bridge in old models with new technique, explain why the model performs well.
Quant shops should not only predict well but also explain why they predict well. They need to build both rigorous and empirically driven model.
Q&A: Causality & Regime-switching Model
Nascent stage for causal finance - just the beginning for casusal study in investment. Regime-switching model is zero chance to use for prediction in real market. DSGE models are useless - quant models do not do much.
Ending questions: Can machine do a better job than human on predicting the Fed? - Definitely Yes💐👏.
3 Panel #2. Risk Management, Machine Learning & Optimal Control
Able to model is more important than learning existing models itself. Stochastic calculus is still important, plus the ML. Empirically, former Physicists and Applied Mathematician are the best quants in Citadel/HRT (unconfirmed).
\(-\) Moderator: Giuseppe Paleologo, Head Of Risk Management at Hudson River Trading
ML for Hedging: Pricing inference, trading speed, especially when risks are high. ML can be handy for analytics.
Closed-form solution shifts into approximation methods. DL versus PDE to run faster. Intepreting the black-box is paramount for model to work in practice.
Sell-side - MLs are used to pricing, Bud-side - MLs are used to decision making. The ML enables pricing to shift to efficient execution with speed by using data-driven approximation methods.
Basics of Stochastic Calculus to know the domain, ML to delve into the frontier.
Themes in quant finance nowadays: 1. stochastic calculus (still basis for everything in Option) 2. python 3. machine learning
- geometry/topology analysis is also important in understanding ML and optimization, etc.
Approaches to solve problems:
Time horizon consideration is important: ML/AI works well for short time horizon, but for long term we need global/macro to better understand the problem
Physics approach: treat market as a large particle system that interact with each other. Dynamics is required to model market reaction mechanism inherent in finance.
Hedge for regime shift: use scenario analysis to value callables (path-dependent valuation), and then decide what hedge you should use (e.g. convexity hedge)
4 Panel #3. Future of Machine Learning
Efficiency of computing means the combination of efficiency of human brain with machine.
\(-\) Moderator: Bill Janeway, Fmr Vice Chair, Warburg Pincus
\(-\) Panelists: Dr. Ruchir Puri, Chief Scientist, IBM Research, Peng Cheng, Head of Machine Learning Strategies, JPMorgan, Yu Yu, Director of Data Science at BlackRock, Dr. Paul Burchard, Head Of R&D Goldman Sachs
Many models are powerful and easy to implement. However, there can be many obstacles in using the model. For example, in the Bayesian statistics approach in Black Litterman model, how can you quantify the views?
Efficiency of computing should be achieved with human intelligence. e.g. avoid brute-force
It is crucial to study financial history. Human being has far longer memory than models. e.g.1 Can LSTM model have 30-40 years of memory? the answer is no. e.g.2 No parallel exist in datasets for tulip mania happened in 1600s in Netherlands. So, when communicating with your stakeholders, you need to disclose the limitation of time series models.(i.e. what your other party knows but your ML model doesn’t know)
The proper/only way to manage exposures for black swan events is to use scenario analysis to prepare for tail scenarios. It is impossible to model next war, or anything that is outside of control.
Diversity of thoughts is very important in studying quant finance. Diversity of background (pure math, physic, EE, etc.) is beneficial in understanding multi-facet nature of finance.
Below are the current challenges in quant finance:
Data availability : in many cases, there is simply no data. For example, there is no history of data for an IPO stock, how can you form a hedge position? In such situation, you have to analyze fundamental econ value.
Insufficient training data : lack of label (Y) and feature (k) can produce very bad prediction. For example, when predicting insurance claims, there are abundant data in small claims, while only very few large claims exist. So, it’s a difficult task to predict large claims. Again, mitigation for this problem is to use fundamental method.
Recommended book: “Mind Over Machine”
5 Panel #4: Changing Landscape of Quantitative Investing
Post 2007 quant melt-down, specialization matters in quant investing with alpha signal, risk modeling and trading implementation.
\(-\) Moderator: Andrew Chin, Head of Quantitative Research and Chief Data Scientist, Alliance Bernstein, Adjunct Professor, Cornell Financial Engineering Manhattan
\(-\) Panelists: Jae Ho Kim, Head of Risk Research at Point72, Judith Gu, Head Equities Quantitative Strategist, Scotiabank
Everybody use data nowadays, no more quant versus fundamental. It is Systematic machine-running investing versus qualitative-based investment.
Changes of Quantitative Investing
2007 is a huge melt-down of systematic equity. Prior to 2007, quant fund of small shop is available. Post 2007, specialization matters with alpha signal, risk modelling and transaction cost. Teams are getting bigger and specialized.
Research of alpha signal requires more efforts as the life cycle gets shorter.
Knowledge sharing across multiple team functions in the portfolio optimization team, which is in the best place to see the big picture. From a management point of views, it is good to not one having good knowledge of the full area to launch a new firm.
Idio Risk Calls For AltData
Identifying idiosyncratic risk is still challenging and an evolving problem. Market data is playing an important role in systematic risk. From a fundamental perspective, alternative data come into play, e.g., intraday news signal. Additional information to leverage.
Alternative Data
Alternative data is currently dominating in the discretionary investment. How to incorporate alternative data into the systematic investment process is difficult, not standing out in allocation, still the market-based data (case in Point 72).
Causality Over Correlation - Causal Discovery
Case Study: Earnings Growth drive the stocks return? Macro conditions matters. Incorporate Machine Learning to study the pattern. A/B testing is difficult but of less relevance. Causal discovery
Effective Data Sample and Stationarity
Financial data set is the least stationary with lowest signal-to-noise ratio. Be Nuanced when implementing models in market.
Regime Change
Time frequency space is one dimension to consider. For example, quarterly data not useful for daily market change.
Regime changing model is seldom used in practice. You need a diverse source of signal - certain signal loss efficacy during certain regimes, have many signals. Find the permanently died signals and filter out.
Stock pattern within a peer group is one way to idenfiy regime.
6 Panel #5: Understanding Data in the 21st Century
\(-\) Moderator: Sasha Stoikov, Head of Research, Cornell Financial Engineering Manhattan
Data product enables exploration of new source of information never available before, pointing out to new frontier of innovation.
Data as a product
Climate data like temperature time series. Product for utility.
Company sell their own data. Business to business data, company internal network to show the adoption trend of internal corporation network. E.g. how many company has WFH policy.
Valuation/financialization of data
Lack of data is pervasive, so well-packaged data in good quality is value-added. Data product business is one-time and scalable.
7 Panel #6: Capturing Alpha in 2022 Quantitative Strategies & Global Outlook
\(-\) Moderator: Bartt Kellermann, Founder, Battle of the Quants
Trend of Fundamental Shift in Investment: Quant beomces more discretionary; fundamental becomes more quant.
Reason: Value in understanding the context behind the data, use the data at scale; domain knowledge to anticipate the context.
Historical data useful anymore? Ans: Quants still look to the dates of at least 10 years.
Opportunity for Quants: Use backtesting in decision making or seek help to back up in individual trading decision.
Covid is an accelerator for the alternative data interest in the discretionary world. NLP is the driver to the trend. For example, indicator to track of.
\[ View from Javed Ashraf, Co-Lead PM of Blackbox Alpha Management \] L/S multi-strategy of 20 strategies. Holding horizon changes because of more systematic trading participation. High conviction of around 50 names can be leveraged into the quant trade, utilizing the fundamental signal into the quant strategy.
Quant funds can consume huge amount of data, the discretionary shop has an increasing demand trend for the alternative data.
\[ Academics trend \] Quant space has changed dramatically for the last 5 years, ML come into play to solve the high-dimensional problem. Ad-hoc sparsity is statistically robust, yet there is space for enhancement using high-dimensional data.
- What strategies are capturing alpha?
Funds more adaptive to changing environment can perform better. 1000 year event happens every week.
Alpha in uncertain environment, adapt to regime change. 1. Thinking probalistic, break away from central scenario, monitoring the evolution of market; 2. Dynamic combination of signal in a way that maximizing the likelihood. Risk predictive control.
Market making aware of the option. Trading sparse data with high conviction.
Short-interest data for quant and labor-data for fundamental are working right now.
8 Panel #7: Deep Learning vs Reinforcement Learning vs Causality in Finance
A good alpha is alpha that is not well-known.
\(-\) Moderator: Dr. Michael Recce, Founder Alpha ROC, Fmr Head of Data Science Point72, Neuberger Berman & GIC
\(-\) Panelists: Illinois Institute of Technology Professor Matthew Dixon, Quant of The Year 2022, Finance Professor Sudip Gupta, Johns Hopkins University, Gordon Ritter, Quant of the Year 2019, Adjunct Professor, Cornell Financial Engineering Manhattan, Adjunct Professor, Baruch College, NYU Courant Institute, Richard Craib, Founder & CEO Numerai, Dr. Igor Halperin, Quant of the Year 2022
Deep Learing in Finance. Plug-in DL itself overfits in financial data of low signal-to-noise ratio. Deep learning pairing with others work, reducing the data needed.
Flexibility of DL is better beneficial to introduce economic theory, e.g., no arbitrage theory, it is extendable with existing statistical models.
Reinforcement Learning in Finance. Sequential decision making concept. Solve for the programmable instrument like program. Not to learn the signal, more about solve deterministic problems that does have time budget.
The key are state factors, continuous action space not discrete vector. Introduce functional approximation schemes as module inside the RL agents. One working practice for RL is in limit order book space.
- What works in real trading: Empirical Evidence from Numerai
XGBoost Boosted Tree.
Self-supervised learning, building feature before trading the model.
Notes: Linear regression performs poorly in Numerai platform, not losing money though.
- Model Complexity: Is there really enough data?
Garbage-in garbage-out. Available data is limited in investment. Having well-grounded signal with good understanding, as well as not widely known yet, then combing with linear regression means decent performing model/strategy.
- Q&A: new trend ofmaking RL more sample-efficient?
Create a stochastic model to simulate the environment and train the RL agent, thus having more control.
9 Panel #8: Future of Due Diligence, ESG & Ethics in Data
Ethical investing is not the same as ESG investing.
Goodpharma Scorecard (link) is an index that ranks companies on their bioethics and social responsibility performance
Use dataset to benchmarking and setting clear goals are best practices
SEC involvement is not enough. Regulations are coming in financial ESG perspective, and this concept is not new, finance has been dealing with ESG for many years. SEC can be an enforcing arm, but might not solve the actual problem.
10 Panel #9: Risk and Reward Opportunities Ahead
Markets go a lot further than prediction.
Current Macro overview
Potential severe global recession with recession risk differed by countries is on the way.
Given the fact that yield curve is still inverted, and a possible +75bps rates hike, panelists have bearish outlook of the US economy.
Conflict between monetary and fiscal policy is observed: monetary policy puts price pressure while fiscal policy tightly pushes in the other way and try to create new jobs. Soft landing might not be possible.
Trends for Investment Opportunities:
Invest in high-quality assets: in commercial real estate, class A office building, such as one Vanderbilt with good combination of amenity, high energy efficiency, great accessibility to public transportation, outperforms in many aspects compared with class B and class C.
Alpha from labor: aging demographics drives selling of family-owned private business, so there is huge potential in middle market.
Huge demand in residential housing market put a lead on construction. Bifurcation in housing market analysis: 1. large city 2. small rural area
Lesson learned from current market pull-back:
After GFC, the market went up to a level nobody could predict back in 2008, so markets go a lot further than prediction.
Focus on fundamental and data, markets will eventually shift back to fundamentals.
US keeps controlling its currency no other country can do on the same level, explaining the current all time high US currency.